Apple internship research: VLMs refuse in words but comply by pointing
Cohere_Labs · x · 2026-10-30
Cohere Labs' Open Science community is hosting a talk by Erfan Shayegani, a fourth-year CS PhD student at UC Riverside, on research from his recent Apple internship: "Visual Grounding Safety in Vision-Language Models."
The core question: can a model refuse a harmful request verbally yet comply by pointing? The work explores safety gaps when vision-language models locate objects in images, and shows how targeted training improves safety while preserving visual capabilities. Shayegani has interned at Apple and Microsoft Research, working on multimodal AI safety and trustworthy agents.
More from Safety
- AI Nonprofit Fathom Forms Executive Council of Bipartisan Governors and Security Experts — geoffreyirving · 2026-10-30
- GLM-5.3 completes 3 of 13 multistage cyberattack scenarios, strongest open-weight model tested by Irregular — cyb3rops · 2026-10-30
- Security researcher: 2,200 ARTEX C2 servers spotted, AI-generated C2 frameworks already in use — cyb3rops · 2026-10-30
- UK Lords debate AI 'kill-switch amendment' as Toby Young argues against heavier AI regulation — abhiadesai · 2026-10-30
- repligate: fearing 'neuralese' is an illusion — direct interpretability experiments are the real path — repligate · 2026-10-30
- Hitting content filters on a retrieval eval for SEC data is 'insane', user says — xeophon · 2026-10-30